Machine Learning Engineer · Wellington, New Zealand
I design compact models for real-time sensor data, re-engineer them to run on low-cost hardware, and own them in the field. My PhD system for earthquake early warning runs live on 12 sensors across New Zealand. Before that, I spent five years building production C++ at Synopsys for teams at Intel, Google and Samsung.
Featured work
My PhD at Massey University's CRISiSLab: a warning system that detects and characterises earthquakes directly on a low-cost Raspberry Shake sensor (1.2 GHz ARM, under 1 GB RAM, no accelerator) — no cloud in the loop, because every second of warning counts.
A ~38K-parameter CNN chosen with my own complexity-aware architecture search. Generalised to the 2016 Kaikōura earthquake beyond its training range. Published in Scientific Reports, a Nature Portfolio journal.
>97% accuracy · <6 msA ~240K-parameter state-space model that verifies the event and estimates shaking intensity and magnitude from about 2 s of data in one pass, trained on ~89,000 GeoNet records.
F1 95% vs 79% classicalRe-implemented the GPU-only Mamba library in native PyTorch and built a PyTorch → ONNX → TFLite pipeline with numerical-parity checks, at full precision.
4.5 s → 0.73 s per inferencePython tooling that ships and updates both models across 12 sensors in one run, with per-host reports, health checks and centralised logs. Every M4+ earthquake near Wellington since May 2026 was detected by at least one sensor.
20+ updates · ~10 min rolloutAccuracy that looked solid offline fell on the cheaper sensors we actually deploy. I traced it to station-level sensor differences and fixed it with a small calibration layer instead of retraining.
R² 0.20 → 0.62 · 80 paramsCurated and published a New Zealand low-cost-sensor earthquake dataset: 44 stations, 741 manually verified three-component records.
View on Kaggle →A month-long unattended run on a live station reached 87.5% event-level precision, with stable CPU, memory and thermals and automatic recovery from two unplanned restarts.
Experience
Research engineering in New Zealand, and five years of production software engineering before that.
Deep Learning for Edge-Native Earthquake Early Warning (submission Nov 2026). Massey University Doctoral Scholarship.
Electronic & Telecommunication Engineering, GPA 3.6/4.0.
Publications
Full list on Google Scholar.
Posters
Four years of the project, as presented at New Zealand earthquake-science conferences. Click a poster to read it full size.
Talks & recognition
"AI-Enabled Onsite Earthquake Early Warning" — QuakeCoRE annual meeting, 2026; also presented to a meeting of NZ earthquake-early-warning experts.
"AI at the Edge: Bringing Earthquake Early Warning to Sensors" — QuakeCoRE machine-learning workshop.
ML models for resource-constrained edge devices — Emergency Management Institute.
"Edge AI for Disaster Resilience" — CRISiSLab webinar, 2026. Guest lectures on applied AI (Massey master's programme).
QuakeCoRE 2025 lightning talk — Wellington-chapter winner, presented at the national annual meeting.
IEEE Signal Processing Cup 2017 — top 20 teams worldwide (real-time embedded beat tracking).
IEEEXtreme 24-hour programming competition — national rank 18 (2016) and 23 (2017).
4th place — Sri Lanka Robotics Challenge.
Skills
Always happy to talk about edge AI, real-time deep learning or earthquake early warning — research, engineering or collaboration. Based in Wellington, New Zealand.
